YuE2: 3.59B-Parameter Music Generation Model Tops SongBench and Arrives on Hugging Face via ComfyUI Repackaging

m-a-p's YuE2 (3.59B parameters, 28 layers) scores 6.9632 on SongBench, beating Suno v5 (6.8721) via symbolic planning, while a bf16 ComfyUI repackaging of YuE2-3B and SheetSage2 trends #30 on Hugging Face.

YuE2 is a music generation model of roughly 3.59B parameters and 28 layers that supports song creation, covering, and agentic editing through semantic music tokens and editable ABC symbolic scores. Its best-of-8 setting reaches 6.9632 on SongBench, the highest mean among 15 evaluated settings on WildSongBench (192 prompts), ahead of Suno v5 at 6.8721. The project also introduces MERT2, whose 632M-parameter encoders achieve state of the art on 14 of 15 MARBLE music-representation metrics, and SheetSage2, which transcribes beats, downbeats, key, chords, structure, and melody with SOTA on 10 of 13 audio-to-score transcription benchmark metrics across six tasks. In a related distribution development, Comfy-Org has published repackaged bf16 safetensors of m-a-p's YuE2-3B model and a sheetsage2_bf16 audio encoder, organized into ComfyUI checkpoints and audio encoder folders and linking to the original m-a-p/YuE2-3B and m-a-p/SheetSage2 model pages; that repository is currently trending #30 on Hugging Face (noted 2026-09-11).

• YuE2 is a music generation model with ~3.59B parameters and 28 layers (also distributed as YuE2-3B), supporting song creation, covering, and agentic editing via semantic music tokens and editable ABC symbolic scores
• Best-of-8 YuE2 scores 6.9632 on SongBench, the highest mean among 15 evaluated settings on WildSongBench (192 prompts), ahead of Suno v5 at 6.8721
• MERT2 encoders (632M parameters) achieve state of the art on 14 of 15 MARBLE music-representation metrics
• SheetSage2 transcribes beats, downbeats, key, chords, structure, and melody, achieving SOTA on 10 of 13 audio-to-score transcription benchmark metrics across six tasks
• Comfy-Org published repackaged bf16 safetensors of m-a-p's YuE2-3B and a sheetsage2_bf16 audio encoder for direct ComfyUI use, linking to the original m-a-p model pages; the repo is trending #30 on Hugging Face